scib.metrics.silhouette_batch

scib.metrics.silhouette_batch(adata, batch_key, group_key, embed, metric='euclidean', return_all=False, scale=True, verbose=True)

Batch ASW

Modified average silhouette width (ASW) of batch

This metric measures the silhouette of a given batch. It assumes that a silhouette width close to 0 represents perfect overlap of the batches, thus the absolute value of the silhouette width is used to measure how well batches are mixed. For all cells \(i\) of a cell type \(C_j\), the batch ASW of that cell type is:

\[batch \, ASW_j = \frac{1}{|C_j|} \sum_{i \in C_j} |silhouette(i)|\]

The final score is the average of the absolute silhouette widths computed per cell type \(M\).

\[batch \, ASW = \frac{1}{|M|} \sum_{i \in M} batch \, ASW_j\]

For a scaled metric (which is the default), the absolute ASW per group is subtracted from 1 before averaging, so that 0 indicates suboptimal label representation and 1 indicates optimal label representation.

\[batch \, ASW_j = \frac{1}{|C_j|} \sum_{i \in C_j} 1 - |silhouette(i)|\]

Parameters

• batch_key – batch labels to be compared against

• group_key – group labels to be subset by e.g. cell type

• embed – name of column in adata.obsm

• metric – see sklearn silhouette score

• scale – if True, scale between 0 and 1

• return_all – if True, return all silhouette scores and label means default False: return average width silhouette (ASW)

• verbose – print silhouette score per group

Returns

Batch ASW (always) Mean silhouette per group in pd.DataFrame (additionally, if return_all=True) Absolute silhouette scores per group label (additionally, if return_all=True)